Model Specific Capability Testing

via “model-capability-reflection-in-training-data”

300K instructions extracted directly from aligned LLM outputs.

Unique: Explicitly designs the data generation process to capture the source model's own capability understanding by having the model generate both instructions and responses. This creates a tight coupling between data distribution and model behavior that is difficult to achieve with human-annotated data.

vs others: More faithful to source model behavior than instruction datasets created by having humans write instructions and the model respond, because both instruction and response generation are controlled by the same model's learned patterns.